Batch Normalization Research Articles

TE-Spikformer:Temporal-enhanced spiking neural network with transformer

The integration of Spiking Neural Networks (Maass, 1997) and Transformers (Vaswani et al., 2017) has significantly enhanced performance in the field, achieving substantial improvements while ensuring energy efficiency. This amalgamation bears significant exploratory significance. This paper aims to enhance the network’s capacity for extracting temporal information from neuromorphic datasets. By leveraging the Spikformer (Zhou et al., 2023b) architecture, we introduce a novel network named TE-Spikformer. Through thorough analysis, we identified an imbalance in the Average Firing Rates(AFR) of neurons in the layer preceding the classification head across temporal steps. We thoroughly investigate the root cause of this issue, attributing it to the limitations of the network’s use of Batch Normalization (BN) (Ioffe and Szegedy, 2015) layer. To address these challenges, we propose a Batch Group Normalization (BGN) layer. While maintaining the stability of temporal characteristics in the data, we also introduce the Spike Spatio-Temporal Attention(SSTA) module to enhance the network’s ability to capture temporal information. To validate the effectiveness of our approach, we conducted multiple experiments using neuromorphic datasets, including DVS-CIFAR10, DVS128 Gesture, and N-Caltech101. The experimental results demonstrate that our algorithm consistently outperforms baseline methods, achieving accuracy rates of 99.30%, 89.60%, and 87.42%, respectively, thereby attaining state-of-the-art performance in the field.

Crop disease diagnosis and prediction using two-stream hybrid convolutional neural networks

Crop diseases significantly impact yield and quality, posing a direct threat to food security. The application of Convolutional Neural Networks (CNN) in crop disease recognition has notably improved diagnosis accuracy and efficiency. This study presents an innovative crop disease classification model based on the VGG-16 network. Enhancements include the incorporation of Batch Normalization (BN) and a novel activation function synergizing with Exponential Linear Units (ELU), improving model convergence speed and accuracy. Additionally, Global Average Pooling (GAP) is integrated to streamline the network architecture, and the InceptionV2 module is introduced to extract leaf disease features from different dimensions, enhancing model robustness. Validation on the PlantVillage dataset shows an accuracy rate of 98.89%, demonstrating the model's competitiveness and its potential to support sustainable agricultural production.

Transient electromagnetic inversion to image the shallow subsurface based on convolutional bidirectional long short-term memory neural networks

SUMMARY The conventional transient electromagnetic inversion method has a low calculation speed and precision and is susceptible to falling into local minima, which does not meet the fine detection requirements of urban underground space. In this study, we proposed a novel inversion method based on convolutional bidirectional long short-term memory neural networks for shallow subsurface transient electromagnetic inversion. This network structure possessed strong spatial feature extraction capabilities and a proficient understanding of sequential data, thereby addressing the issues of slow conventional inversion computations and inadequate inversion accuracy. Utilizing the apparent resistivity from a three-layer model as the sample input and the real model as the target, the network was trained using batch normalization and dropout techniques to accelerate the convergence rate. The resulting model achieved real-time inversion speeds and high accuracy, with robust generalization capabilities and adaptability to new data. To assess the inversion performance, we used a novel 1-D inversion error calculation index, the correlation area loss error, for a more accurate measurement. Numerical simulation experiments showed that the proposed method required only 2.121 s to invert data from 100 observation points. The inversion efficiency was significantly superior to the conventional methods, maintaining excellent accuracy while effectively discerning subsurface electrical stratification in geophysics. Applying convolutional bidirectional long short-term memory neural networks to multidimensional and field data yielded results superior to those of conventional inversion, demonstrating the promising applicability and generalization of this approach. This study offers an efficient solution for shallow subsurface transient electromagnetic exploration and holds potential for application in other areas.

Diseño de una aplicación para detectar covid-19 mediante redes neuronales convolucionales e imágenes de rayos X

This research presents the design of an application to detect covid-19 using convolutional neural networks and X-ray images in two scenarios (covid/Non-covid and covid/Normal/Pneumonia). To avoid overfitting online data augmentation, dropout, batch normalization, and Adam optimizer was used. The three-class network was used as a pre-trained model, tuning only the dense and output layers to obtain the binary model. Additionally, hyper-parameter optimization was used to get dropout probabilities, activation functions, and neurons. The learning rate was adjusted using callbacks to avoid local optimums. Networks were converted to TensorFlow.js format and embedded locally in a hybrid application using Ionic and Capacitor and were deployed through Firebase to help provide diagnostics. The application obtained an accuracy of 98.61% and 96.48% for two and three classes, respectively, achieving higher performance when compared to other proposals (offline models) in the literature and using fewer training parameters.

Optimizing Data Flow in Binary Neural Networks.

Binary neural networks (BNNs) can substantially accelerate a neural network's inference time by substituting its costly floating-point arithmetic with bit-wise operations. Nevertheless, state-of-the-art approaches reduce the efficiency of the data flow in the BNN layers by introducing intermediate conversions from 1 to 16/32 bits. We propose a novel training scheme, denoted as BNN-Clip, that can increase the parallelism and data flow of the BNN pipeline; specifically, we introduce a clipping block that reduces the data width from 32 bits to 8. Furthermore, we decrease the internal accumulator size of a binary layer, usually kept using 32 bits to prevent data overflow, with no accuracy loss. Moreover, we propose an optimization of the batch normalization layer that reduces latency and simplifies deployment. Finally, we present an optimized implementation of the binary direct convolution for ARM NEON instruction sets. Our experiments show a consistent inference latency speed-up (up to 1.3 and 2.4× compared to two state-of-the-art BNN frameworks) while reaching an accuracy comparable with state-of-the-art approaches on datasets like CIFAR-10, SVHN, and ImageNet.

A stable and robust fault diagnosis method for bearing using lightweight batch normalization-free residual network

Abstract The conventional deep learning-based bearing fault diagnosis methods tend to utilize denoising modules to improve the fault diagnosis performance in noisy scenes. However, the addition of denoising modules will increase expensive computational costs, leading to a delayed acquisition of fault diagnosis results. This work proposed a lightweight batch normalization (BN)-free residual network without any denoising modules for bearing fault diagnosis which properly rescaled the weights in a standard initialization instead of BN to avoid the exploding gradient problem and vanishing gradient problem at the beginning of training for deep neural networks. Therefore, it prevents the undesirable properties caused by BN. Compared with other methods, the fault diagnosis performance of the proposed method can maintain a high level with different input sizes and batch sizes. Especially in noisy scenes, the testing accuracy of fault diagnosis on different bearing datasets can be improved by 13.54% and 7.74% using fewer parameters and floating point operations on different bearing datasets.

Domain adaptation using AdaBN and AdaIN for high-resolution IVD mesh reconstruction from clinical MRI.

Deep learning has firmly established its dominance in medical imaging applications. However, careful consideration must be exercised when transitioning a trained source model to adapt to an entirely distinct environment that deviates significantly from the training set. The majority of the efforts to mitigate this issue have predominantly focused on classification and segmentation tasks. In this work, we perform a domain adaptation of a trained source model to reconstruct high-resolution intervertebral disc meshes from low-resolution MRI. To address the outlined challenges, we use MRI2Mesh as the shape reconstruction network. It incorporates three major modules: image encoder, mesh deformation, and cross-level feature fusion. This feature fusion module is used to encapsulate local and global disc features. We evaluate two major domain adaptation techniques: adaptive batch normalization (AdaBN) and adaptive instance normalization (AdaIN) for the task of shape reconstruction. Experiments conducted on distinct datasets, including data from different populations, machines, and test sites demonstrate the effectiveness of MRI2Mesh for domain adaptation. MRI2Mesh achieved up to a 14% decrease in Hausdorff distance (HD) and a 19% decrease in the point-to-surface (P2S) metric for both AdaBN and AdaIN experiments, indicating improved performance. MRI2Mesh has demonstrated consistent superiority to the state-of-the-art Voxel2Mesh network across a diverse range of datasets, populations, and scanning protocols, highlighting its versatility. Additionally, AdaBN has emerged as a robust method compared to the AdaIN technique. Further experiments show that MRI2Mesh, when combined with AdaBN, holds immense promise for enhancing the precision of anatomical shape reconstruction in domain adaptation.

Deep Learning–Assisted Parameter Monitoring and Optimization in Rotary-Percussive Drilling

Summary As an efficient method for hard rock fracturing, rotary-percussive drilling has been widely used in various scenarios, especially deep drilling. Drilling parameter monitoring and control are necessary to ensure stable and efficient underground drilling processes. However, this may be more difficult in deep, harsh conditions. In this paper, our goal is to establish models based on deep learning for drilling parameter monitoring and optimization. Combining impregnated diamond bits and granite rock samples, we conducted rotary-percussive rock drilling experiments using a rock drilling test rig. Real-time acoustic signals during rotary-percussive drilling were recorded, segmented, and transformed as spectra, which made up a drilling acoustic signal data set. Drilling parameters, including rotational speed (revolutions per minute, RPM), pump flow rate, pump pressure, weight on bit (WOB), torque, and rate of penetration (ROP), were logged in the meantime. Given the acoustic signal as input, we built 1D convolutional neural network (1D-CNN) models for drilling parameter prediction. The prediction results revealed the high efficiency and accuracy of 1D-CNN regression models based on deep learning in drilling condition monitoring. Batch normalization played an essential role in the regression model training processes. Given that these parameters have different units and dimensions, we compared models with different output modes to evaluate the multiparameter prediction performance of the 1D-CNN. Taking RPM, flow rate, pressure, and WOB as independent variables and torque and ROP as dependent variables, we developed a conditional variational autoencoder to realize optimization on drilling parameters based on expected drilling performance.

Copyright protection framework for federated learning models against collusion attacks

Federated learning (FL) models are constructed by multiple participants who provide their training datasets and collaborate in joint training. However, training and deployment processes have encountered various challenges in terms of intellectual property protection, such as illegal theft and data leakage. Existing FL protection frameworks focus on each client independently and verify the model ownership. When they are under collusion attacks (i.e., when multiple clients negotiate to steal together), they cannot accurately identify the clients that have stolen the model. To address this challenge, a novel watermarking protection scheme against collusion attacks for federated learning is proposed in this work. It employs anti-collusion coding to design unique watermark information for each client, which can effectively detect colluders. Furthermore, it utilizes a specific regularized loss function for watermark information embedding along with the incorporation of skip connections to embed the watermark information within each batch normalization layer. The experimental results demonstrated that embedding different watermark information into each client did not affect the accuracy of the original task. The accuracy was approximately 100% when identifying the colluders. The embedding and extraction times for the original task were only 1.53% and 0.29%, respectively. Further, it exhibited high robustness against various common attacks, including fine-tuning, shearing, and collusion attacks.

All signals point to personality: A dual-pipeline LSTM-attention and symbolic dynamics framework for predicting personality traits from Bio-Electrical signals

The prediction of personality traits offers valuable insights into human behaviour, more specifically in psychology, healthcare, and social science. In this paper, we present a novel methodology for personality trait prediction using a dual-pipeline architecture. The model architecture leverages Long Short-Term Memory (LSTM) networks with batch normalization for capturing sequential dependencies in data and incorporates temporal attention heads for feature extraction. By combining these parallel pipelines, our network effectively utilizes both LSTM and attention mechanisms to create a comprehensive representation of input data. The network’s goal is to predict the OCEAN (openness, conscientiousness, extraversion, agreeableness and neuroticism) traits using physiological signals including: EEG, ECG and GSR. Including attention mechanisms enables the model to focus on critical moments in these signals, resulting in significantly improved prediction accuracy. Experimental evaluations demonstrate the superior performance of our method compared to traditional machine learning methods on two publicly available datasets: ASCERTAIN and AMIGOS. Our source code is accessible at https://github.com/deepakkumar-iitr/AT3NET.

Hybrid attention transformer with re-parameterized large kernel convolution for image super-resolution

Single image super-resolution is a well-established low-level vision task that aims to reconstruct high-resolution images from low-resolution images. Methods based on Transformer have shown remarkable success and achieved outstanding performance in SISR tasks. While Transformer effectively models global information, it is less effective at capturing high frequencies such as stripes that primarily provide local information. Additionally, it has the potential to further enhance the capture of global information. To tackle this, we propose a novel Large Kernel Hybrid Attention Transformer using re-parameterization. It combines different kernel sizes and different steps re-parameterized convolution layers with Transformer to effectively capture global and local information to learn comprehensive features with low-frequency and high-frequency information. Moreover, in order to solve the problem of using batch normalization layer to introduce artifacts in SISR, we propose a new training strategy which is fusing convolution layer and batch normalization layer after certain training epochs. This strategy can enjoy the acceleration convergence effect of batch normalization layer in training and effectively eliminate the problem of artifacts in the inference stage. For re-parameterization of multiple parallel branch convolution layers, adopting this strategy can further reduce the amount of calculation of training. By coupling these core improvements, our LKHAT achieves state-of-the-art performance for single image super-resolution task.

Insulation life prediction of stacking multi-model fusion motors based on Conv1d-LSTM

Abstract The insulation system of the mine motor is key to ensuring its stable operation, which affects the efficiency of mine production and transportation and the benefit of enterprises. The prediction of motor insulation life can effectively prevent the deterioration of motor performance. Due to the complex underground environment, many factors affect the insulation of coal mine motors, which are not conducive to the selection of feature parameters. The insulation life prediction efficiency and accuracy of motors need to be improved. To solve this problem, this paper proposes a prediction method based on Conv1d (one-dimensional convolve) and LSTM (long short-term memory neural network), which integrates multiple models and the Stacking integrated algorithms. The operating data and environmental parameters of a coal mine motor under different conditions are analyzed. Three Conv1d-LSTM hybrid models with batch normalization, residual network, and dense number are used as the base learner, and the results are weighted by K-fold to verify the accuracy of the three models. The linear regression model is used as a meta-learner to prevent overfitting of the model and effectively improve the speed and accuracy of single network optimization. Compared with the traditional method, the results show that the optimization of the LSTM neural network by the multi-feature model has a higher accuracy in predicting the remaining life of motor insulation, which provides a new way for the prediction of motor insulation life.

A domain adaptation‐based convolutional neural network incorporating data augmentation for power system dynamic security assessment

AbstractRecently, deep learning (DL) based dynamic security assessment (DSA) methods have been very successful. However, although a DSA model can be trained well for a specific topology, it often does not perform well for other topologies. Since the topology in real‐world power systems is frequently changing, the performance reduction of DL‐based DSA methods is very serious, which is a challenging and urgent problem. This paper proposes a novel DSA method based on a convolutional neural network (CNN) to solve this problem. In the proposed method, a strong yet simple domain adaptation approach named adaptive batch normalization (AdaBN) is used, which significantly enhances the extensibility and generalizability of the DSA model when the topology changes and eliminates the need to train a large number of models. This approach achieves a deep adaptation effect by modulating the statistics from the source domain to the target domain in all batch normalization layers across the model. Unlike other domain adaptation methods, this method is parameter‐free, requires no additional components, and has advanced performance despite its simplicity. In addition, this paper introduces TGAN‐based data augmentation to deal with the difficulty of costly data collection and labelling. This data augmentation makes the proposed model applicable to small databases. The test results of the proposed method on IEEE 39‐bus and IEEE 118‐bus systems show that this method can evaluate system dynamic security during topology changes and in the face of data noise with high accuracy.

An advanced heat source localization technology for intelligent warehousing: A multi-source fusion image segmentation approach leveraging infrared and visible light data

Infrared thermography technology, leveraging its unique ability to capture temperature features, has significantly improved the precision of high-temperature target localization. However, infrared imaging technology is limited by issues such as low image contrast, difficulty in distinguishing object categories, and limited image clarity. To enable intelligent detection of high-temperature objects that may cause fires in warehouses, this paper proposes an innovative method that integrates deep learning image segmentation with infrared and visible light image technology. We developed a new image segmentation model based on improved Fully Convolutional Networks and Deconvolutional Networks, introducing a batch normalization layer to accelerate convergence and employing the PReLU activation function to prevent neuron death, thereby enhancing convergence speed and accuracy. Through a feature dynamic image registration method combining a joint model and a cross-modulation strategy, we achieved efficient image fusion. In addition, a game theory-based strategy was adopted to correct localization results, ensuring accuracy. Experimental results demonstrate that the improved model achieves localization accuracy and precision rates of up to 89.30% and 88.00%, respectively, in real-world warehouse heat source scenarios, representing a significant improvement of 9.90% and 2.85% compared to the pre-improvement model, fully validating its advancement and effectiveness.

A SAR Ship Detection Method Based on Adversarial Training.

SAR (synthetic aperture radar) ship detection is a hot topic due to the breadth of its application. However, limited by the volume of the SAR image, the generalization ability of the detector is low, which makes it difficult to adapt to new scenes. Although many data augmentation methods-for example, clipping, pasting, and mixing-are used, the accuracy is improved little. In order to solve this problem, the adversarial training is used for data generation in this paper. Perturbation is added to the SAR image to generate new samples for training, and it can make the detector learn more abundant features and promote the robustness of the detector. By separating batch normalization between clean samples and disturbed images, the performance degradation on clean samples is avoided. By simultaneously perturbing and selecting large losses of classification and location, it can keep the detector adaptable to more confrontational samples. The optimization efficiency and results are improved through K-step average perturbation and one-step gradient descent. The experiments on different detectors show that the proposed method achieves 8%, 10%, and 17% AP (Average Precision) improvement on the SSDD, SAR-Ship-Dataset, and AIR-SARShip, compared to the traditional data augmentation methods.

VashaNet: An automated system for recognizing handwritten Bangla basic characters using deep convolutional neural network

Automated character recognition is currently highly popular due to its wide range of applications. Bengali handwritten character recognition (BHCR) is an extremely difficult issue because of the nature of the script. Very few handwritten character recognition (HCR) models are capable of accurately classifying all different sorts of Bangla characters. Recently, image recognition, video analytics, and natural language processing have all found great success using convolutional neural network (CNN) due to its ability to extract and classify features in novel ways. In this paper, we introduce a VashaNet model for recognizing Bangla handwritten basic characters. The suggested VashaNet model employs a 26-layer deep convolutional neural network (DCNN) architecture consisting of nine convolutional layers, six max pooling layers, two dropout layers, five batch normalization layers, one flattening layer, two dense layers, and one output layer. The experiment was performed over 2 datasets consisting of a primary dataset of 5750 images, CMATERdb 3.1.2 for the purpose of training and evaluating the model. The suggested character recognition model worked very well, with test accuracy rates of 94.60% for the primary dataset, 94.43% for CMATERdb 3.1.2 dataset. These remarkable outcomes demonstrate that the proposed VashaNet outperforms other existing methods and offers improved suitability in different character recognition tasks. The proposed approach is a viable candidate for the high efficient practical automatic BHCR system. The proposed approach is a more powerful candidate for the development of an automatic BHCR system for use in practical settings.

An Efficient Hypertuned DNN Based Approach for Pneumonia Detection

Objectives: To create a deep learning system for pneumonia detection that is both effective and gradually optimized. Methods: A customized CNN is used with an incremental approach for pneumonia classification and detection. Starting with a baseline model, hypertuning parameters such as four convolution layers with filters of 16, 32, 64, and 128 sizes, a dropout layer with values of 0.3, 0.5, and 0.7, four batch normalization layers, and an Adam optimizer are added. A total images of 5,863 for training, 624 for testing, and 16 for validation from the Paul Mooney dataset were used to test the suggested model. Findings: The study recorded a test accuracy of 94% for the customized CNN followed by ResNet50 at 79.9%, VGG16 at 90.14%, VGG19 at 82.21%, InceptionV3 at 74.51%, and EfficientNetB1 at 83.17%. Recall of 98.20%, accuracy of 85.55%, AUC of 93.52%, and F1_score of 92.45% obtained were all fairly excellent for the customized CNN. 15 epochs, a learning rate of 0.0001, callbacks with a patience of 3, and an early stopping feature were applied to the training model. Novelty: Five convolution blocks, two separable convolution layers, one batch normalization layer, one maxpooling layer, and a fully connected layer with an Adam optimizer were all included in the customized CNN that was developed to identify and categorize pneumonia. With Explainable AI's GradCAM technology, pneumonia-infected areas on chest X-rays were highlighted and the sickness was seen. Keywords: Customized CNN, VGG16, VGG19, ResNet50, Explainable AI

Research on Internal Damage Identification of Wire Rope Based on Improved VGG Network.

In order to solve the problem of great difficulty in detecting the internal damage of wire rope, this paper proposes a method to improve the VGG model to identify the internal damage of wire rope. The short-time Fourier transform method is used to transform the wire rope damage signal into a time-frequency spectrogram as the model input, and then the traditional VGG model is improved from three aspects: firstly, the attention mechanism module is introduced to increase the effective feature weights, which effectively improves the recognition accuracy; and then, the batch normalization layer is added to carry out a uniform normalization of the data, so as to make the model easier to converge. At the same time, the pooling layer and the fully connected layer are improved to solve the redundancy problem of the traditional VGG network model, which makes the model structure more lightweight, greatly saves the computational cost, shortens the training time, and finally adopts the joint-sample uniformly distributed cross-entropy as the loss function to solve the overfitting problem and further improve the recognition rate. The experimental results show that the improved VGG model has an identification accuracy of up to 98.84% for the internal damage spectrogram of the wire rope, which shows a good identification ability. Not only that, but the model is also superior, with less time-consuming training and stronger generalization ability.

Normalization Layer Enhancement in Convolutional Neural Network for Parking Space Classification

The research problem of this study is the urgent need for real-time parking availability information to assist drivers in quickly and accurately locating available parking spaces, aiming to improve upon the accuracy not achieved by previous studies. The objective of this research is to enhance the classification accuracy of parking spaces using a Convolutional Neural Network (CNN) model, specifically by integrating an effective normalizing function into the CNN architecture. The research method employed involves the application of four distinct normalizing functions to the EfficientParkingNet, a tailored CNN architecture designed for the precise classification of parking spaces. The results indicate that the EfficientParkingNet model, when equipped with the Group Normalization function, outperforms other models using Batch Normalization, Inter-Channel Local Response Normalization, and Intra-Channel Local Response Normalization in terms of classification accuracy. Furthermore, it surpasses other similar CNN models such as mAlexnet, you only look once (Yolo)+mobilenet, and CarNet in the same classification task. This demonstrates that EfficientParkingNet with Group Normalization significantly enhances parking space classification, thus providing drivers with more reliable and accurate parking availability information.

Data-driven crop growth simulation on time-varying generated images using multi-conditional generative adversarial networks

BackgroundImage-based crop growth modeling can substantially contribute to precision agriculture by revealing spatial crop development over time, which allows an early and location-specific estimation of relevant future plant traits, such as leaf area or biomass. A prerequisite for realistic and sharp crop image generation is the integration of multiple growth-influencing conditions in a model, such as an image of an initial growth stage, the associated growth time, and further information about the field treatment. While image-based models provide more flexibility for crop growth modeling than process-based models, there is still a significant research gap in the comprehensive integration of various growth-influencing conditions. Further exploration and investigation are needed to address this gap.MethodsWe present a two-stage framework consisting first of an image generation model and second of a growth estimation model, independently trained. The image generation model is a conditional Wasserstein generative adversarial network (CWGAN). In the generator of this model, conditional batch normalization (CBN) is used to integrate conditions of different types along with the input image. This allows the model to generate time-varying artificial images dependent on multiple influencing factors. These images are used by the second part of the framework for plant phenotyping by deriving plant-specific traits and comparing them with those of non-artificial (real) reference images. In addition, image quality is evaluated using multi-scale structural similarity (MS-SSIM), learned perceptual image patch similarity (LPIPS), and Fréchet inception distance (FID). During inference, the framework allows image generation for any combination of conditions used in training; we call this generation data-driven crop growth simulation.ResultsExperiments are performed on three datasets of different complexity. These datasets include the laboratory plant Arabidopsis thaliana (Arabidopsis) and crops grown under real field conditions, namely cauliflower (GrowliFlower) and crop mixtures consisting of faba bean and spring wheat (MixedCrop). In all cases, the framework allows realistic, sharp image generations with a slight loss of quality from short-term to long-term predictions. For MixedCrop grown under varying treatments (different cultivars, sowing densities), the results show that adding these treatment information increases the generation quality and phenotyping accuracy measured by the estimated biomass. Simulations of varying growth-influencing conditions performed with the trained framework provide valuable insights into how such factors relate to crop appearances, which is particularly useful in complex, less explored crop mixture systems. Further results show that adding process-based simulated biomass as a condition increases the accuracy of the derived phenotypic traits from the predicted images. This demonstrates the potential of our framework to serve as an interface between a data-driven and a process-based crop growth model.ConclusionThe realistic generation and simulation of future plant appearances is adequately feasible by multi-conditional CWGAN. The presented framework complements process-based models and overcomes their limitations, such as the reliance on assumptions and the low exact field-localization specificity, by realistic visualizations of the spatial crop development that directly lead to a high explainability of the model predictions.